On 12/14/2015 07:09 PM, Paolo Bonzini wrote:
On 14/12/2015 16:33, Andrey Smetanin wrote:
It's possible that guest send us Hyper-V EOM at the middle
of Hyper-V SynIC timer running, so we start processing of Hyper-V
SynIC timers in vcpu context and stop the Hyper-V SynIC timer
uncoditionally and lose time expiration which Windows 2012R2 guest
expects.
The patch fixes such situation by not stopping Hyper-V SynIC timer
at all, because it's safe to restart it without stop in vcpu context
and timer callback always returns HRTIMER_NORESTART.
Can you summarize with a "picture" what is the bad race?
Currently I see that guest starts periodic timer and doesn't clear
message slot after timer expires, so timer expires again and trying to
deliver expiration message but message slot is still busy so we set
->msg_pending flag for guest to receive EOM. timer restarts again and
while it's not expired guest notifies us with EOM, in this case we
schedule timer processing in vcpu context by KVM_REQ_HV_STIMER,
kvm_hv_process_stimers() is called in vcpu context and stops the timer
before it expires, so timer is disabled forever but guest expects it's
periodic expiration(15ms).
I do not understand why Windows doesn't clear message slot for a long
time, it's likely need to be analyzed with debugger(and need more
research). But we can go out from such situation by such fix.
The patch seems safe, but I'd like to have a better understanding of
what goes wrong.
Paolo
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
